NEW YORK – Researchers from the Novartis Institutes for BioMedical Research (NIBR) have developed a method to determine clonal hematopoiesis in cancer patients from cell-free DNA (cfDNA) alone, using a computational model that can discriminate between blood-derived mutations arising from clonal hematopoiesis of indeterminate potential (CHIP) and tumor-derived mutations.
CfDNA sequencing is gaining wide usage in clinical trials as a noninvasive approach for determining the genomic landscape of cancer, monitoring minimal residual disease, and potentially detecting cancer early.
However, up until now, it has been difficult to tell apart mutations that derive from blood cells, potentially from CHIP, and those that come from solid tumors without sequencing DNA from matched white blood cells along with the cfDNA.
"Unfortunately, WBC sequencing isn't done routinely. One of the reasons is that it is expensive," said corresponding author Lauren Fairchild, a data scientist at NIBR.
Her team's new computational method, published in Science Translational Medicine on Wednesday, doesn't need WBC sequencing data to determine the presence of clonal hematopoiesis from cfDNA. "Our open-source method provides a computational solution to labeling CH variants from cfDNA, reducing the need for costly matched WBC sequencing in the clinical setting," the authors wrote.
The method was developed using a published training dataset containing tumor, plasma, and matched WBC sequencing data from 124 patients with metastatic cancer and 47 healthy controls. The tumor data were generated with Memorial Sloan Kettering's MSK-IMPACT panel, which targets 410 cancer-associated genes, while the plasma and WBC sequencing data were generated with a panel of 508 genes at cancer early detection firm Grail.
The researchers then selected 1,400 single-nucleotide variants from this dataset that were known to derive from either tumors or matched WBCs in order to train machine learning models to classify the source of variants using just cell-free plasma DNA.
Next, the team applied the model to 4,324 plasma samples from patients with advanced metastatic cancer, including ER+ breast cancer, cutaneous melanoma, non-small cell lung cancer, and colorectal cancer, who were enrolled in clinical trials at Novartis. Data for these samples were generated using two versions of a sequencing panel of about 564 genes.
They found that 30 percent of these patients had variants associated with CHIP, with some tumor types showing a higher incidence of clonal hematopoiesis than others, and CH generally increasing with age.
While previous studies have linked CHIP to poorer overall survival in cancer patients, the new study for the first time demonstrated that patients with CHIP also showed high inflammatory markers, both systemically and in their tumor microenvironment, in particular an increase in neutrophils.
"These findings suggest that CH status as an additional biomarker of [tumor microenvironment] inflammation and its impact on patients’ response to therapy merit further investigation," the authors wrote.
The exact mechanism linking CHIP and increased inflammatory signals in tumors is still a mystery, however, and was one of the limitations of this study, Fairchild said.
Meanwhile, the researchers highlighted that they need to include more cancer types in the machine learning models to improve its accuracy. "For example, patients with melanoma were not present in the published study used as a training dataset. Therefore, some known oncogenic variants in melanoma such as BRAF V600E were not as consistently classified as tumor-derived as common oncogenic variants in represented cancer types, such as those in EGFR and KRAS," the authors wrote.